Open Source Overview - Negentro Docs

Run PiyAPI on your own infrastructure with full control over your data, configuration, and deployment environment. PiyAPI Open Source provides the flexibility to deploy our cognitive memory engine entirely within your VPC. Maintain strict data privacy, enterprise compliance, and total configuration control without sacrificing the power of bitemporal reasoning. Developers can install the SDK directly into an application as a lightweight library, or run the full PiyAPI server as an independent, scalable microservice. Both approaches allow you to configure storage, embeddings, and BYOK (Bring Your Own Key) model routing to match your existing infrastructure. The stack is designed to be fully extensible: connect your existing PostgreSQL databases, plug in custom LLMs, and build active inference workflows tailored to your exact deployment constraints.

Run PiyAPI your way

As a library

Install the lightweight SDK directly into your TypeScript or Python application for immediate, embedded bitemporal memory.

As a self-hosted server

Run the full PiyAPI Gateway and Memory Engine as an independent Docker service with complete control over Redis caching, pgvector indexing, and LLM routing.

Get started

Python Quickstart

Install the piyapi-memory package and verify that hybrid memory storage and retrieval work in a few lines of code.

Node.js Quickstart

Install the @piyapi/sdk package and connect your first stateful agent workflow.

Self-hosted server

Deploy the PiyAPI Docker container locally or on your Kubernetes cluster and configure your bitemporal infrastructure.

Go further

Configure components

Configure BYOK routing, embedding providers, pgvector storage, and tune the UnifiedScorer for your specific deployment.

Self-hosting features

Explore enterprise namespace isolation, semantic cache purging, multimodal CDC connectors, and compliance-grade redaction tools.

Build with cookbooks

Use practical examples for LangChain agents, CrewAI, MCP server integrations, and real-world compliance workflows.

Need a managed alternative? Compare the self-hosted open-source deployment with the fully managed PiyAPI Cloud Platform, or switch to the Platform documentation.

Default components

PiyAPI ships with sane, production-ready defaults for each part of the pipeline. Every component can be replaced or explicitly configured to match your existing VPC infrastructure and provider preferences.

Library defaults

Component	Default
LLM	OpenAI (via BYOK)
Embeddings	OpenAI `text-embedding-3-small`
Vector store	`pgvector`
History store	PostgreSQL

Self-hosted defaults

Component	Default
Reranker	UnifiedScorer (BM25 + RRF)
Storage	PostgreSQL (Bitemporal KG)
Metadata cache	Redis
Providers	Multi-Provider BYOK Routing

Configure your stack

// piyapi.config.ts
import { createPiyAPI } from '@piyapi/sdk';

const piyapi = createPiyAPI({
  llm: "openai:gpt-4o",
  embeddings: "openai:text-embedding-3-small",
  // Connect directly to your existing infrastructure
  vectorStore: "pgvector://localhost:5432/piyapi",
  storage: "postgres://localhost:5432/piyapi_graph",
  cache: "redis://localhost:6379"
});

Self-hosted architecture

[1] Your Application -> [2] PiyAPI Server -> (Memory, Retrieval, Embeddings, LLM, Storage) -> [3] Your Infrastructure

Configuration

LLMs

Route generation through any provider using our unified BYOK manager. Support for OpenAI, Anthropic, Gemini, DeepSeek, Mistral, Groq, Cohere, and Perplexity is handled natively.

Embeddings

Choose the embedding provider and model used for vector generation. Easily swap out defaults for specialized models depending on your latency and accuracy requirements.

Vector databases

Connect the storage backend used for dense HNSW retrieval. Defaults to PostgreSQL with pgvector for robust, transaction-safe vector operations.

Storage

Configure where persistent bitemporal graph data (PiyGraph) and conversational history are stored.

Metadata filtering

Control how stored context is filtered and retrieved using strict namespace scoping (X-Namespace-Prefix) and arbitrary tagging rules.

Reranking

Fine-tune the UnifiedScorer reranking layer (combining BM25 trigram search and 11 distinct retrieval signals) when hyper-precise retrieval is required for your agents.

Build on the open-source stack

Replace infrastructure components (for example, swap Redis for an alternative caching layer)
Configure multi-provider routing and failover logic to prevent LLM downtime
Integrate directly with your existing enterprise PostgreSQL databases
Build custom Active Inference workflows and agent MCP tools

Open-source note Keep your deployment configuration explicit so model, storage, and retrieval dependencies remain easy to understand, version control, and operate across your team.

​Run PiyAPI your way

As a library

As a self-hosted server

​Get started

Python Quickstart

Node.js Quickstart

Self-hosted server

​Go further

Configure components

Self-hosting features

Build with cookbooks

​Default components

​Library defaults

​Self-hosted defaults

​Configure your stack

​Self-hosted architecture

​Configuration

​Build on the open-source stack

Run PiyAPI your way

Get started

Go further

Default components

Library defaults

Self-hosted defaults

Configure your stack

Self-hosted architecture

Configuration

Build on the open-source stack